Comment by xoa | Hacker News Reader

Comment by xoa | Hacker News Reader

Parent Full thread

All the major manufacturers have had significant issues at one point or another particularly with consumer drives, and it seems to go through cycles as well so there is no satisfying perfect answer here imo. I also haven't seen as much hard data on SSDs and they've been evolving so fast, so everything feels a lot murkier. At least with spinning rust we have Backblaze's drive stats going back many years now.

- If it's a single drive situation and reliability is very important, something aimed at "enterprise" tends to be a good bet. More firmware testing, more over provisioning, and so on. And if nothing else there tends to be less BS around getting warranty coverage honored. Also costs significantly more. Micron, Intel D3s etc.

- If it's multi-drive, don't focus on finding the one best brand, use ones from different major brands with redundancy. The goal isn't to try to ensure no drive failures, but rather lower the risk of correlated drive failures so that when they happen the system has time to restore redundancy and there's no issue.

- If it's single drive and some down time is acceptable for cost then just roll the dice, though even there I'd stay away from QLC and the ultra mega bargain basement stuff.

In all cases buy from somewhere reputable that actually sells drives directly, like B&H or CDW or the like. If you buy from Amazon beware of counterfeits and also "new" stuff that is heavily used. Always check SMART stats right away and make sure drive lifetime is where you expect (not hundreds/thousands of hours on something "new"). And of course, always have good backups and/or some way to rapidly bring stuff back up.

Why stay away from QLC? Does it not have its place in high-read low-write consumer workloads, such as software or media libraries?

>Why stay away from QLC?

Same basic idea as SMR hard drives: the savings are too minor to be worth the asterisks. If you're running at megascale and can afford to customize your workloads, software, OS, and firmware, around it and at scale the ROI crunches out as worth it that's fine of course. But at the individual or even SMB level? Nah. Like, taking Samsung as a fairly large, decent brand example, their 4TB TLC EVO drive is ~$220 right now, whereas their 4TB QLC QVO is... $200. Heck for whatever reason their 1TB at B&H costs (a very little) more, $65 for the QVO, while the EVO is $60. Is $5-20 over terabytes and years worth the slightest hassle or weirdness? QLC is IMO essentially a scam at the consumer level, dramatically impacting reliability/flexibility/performance consistency for really minimal cost savings.

>Does it not have its place in high-read low-write consumer workloads, such as software or media libraries?

Frankly no. I mean, putting aside whether SSDs themselves make sense in such an application at all vs spinning rust (same money will buy you a nice 14TB hard drive), but to the extent they do I still think the above holds. QLC is just physically fundamentally sketchier tech. And how sure are you (and how much trouble do you want to go through) about what your workload will be like long term? If you put the money/time investing into getting together a decent little SSD NAS with a bunch of terabytes, and then a year or two down the road want to have some VMs on it too, or have a fileshare and then it gets hit harder than you expected, or there's some bug that causes some write amplification, and then your QLC drives end up hosed was that really worth a few cups of coffee?

Maybe so! SMR really is worth it to a few places as well. Or ZFS dedup for a software example (with hardware implications). But all of this for me falls under the heading of "if you don't have a specific reason for it, don't".

Reply on news.ycombinator.com